PhD student Nathan Whitmore has repurposed a "Too dangerous to release" AI text generator to create a choose your own adventure game.

Whitmore built the text-based game, called GPT Adventure, using OpenAI's GPT-2 language model, which generates convincing text from prompts.

While OpenAI trained its model on 40GB of text from 8 million web pages, Whitmore trained his game on transcripts from 1970s text-based games like Zork and Colossal Cave Adventure, according to Digital Trends.

The original GPT-2 model produced fake news articles and stories so convincing that the company worried it would be misused. Whitmore, however, saw potential for creating exploratory experiences:

You can feed in the player’s current location and their action — [for example], ‘go east’ — and then you just ask GPT-2 to predict what text comes next, which is the consequence of that action. Then the player makes another action, and you have a game.

The game responds to simple prompts from the player and explains what happens, or what the player sees, after they decide what action to take. From Whitmore's blog:

What’s the goal of this game? Exploring, I guess. Even the game is making it up as it goes along, and that’s what makes it interesting.